def test_trivial_solutions_non_unitary(self):
map_computer = MeanAveragePrecisionMetric(0.0, 'non-unitary')
map_computer.update([[0, 0, 26, 26, 1, 0]], [[5, 2, 15, 9, 0],
[18, 10, 26, 15, 0],
[18, 10, 26, 15, 0],
[18, 10, 26, 15, 0],
[18, 10, 26, 15, 0]])
_ = map_computer.compute()
assert np.allclose(map_computer.precision_per_class[0], 1.0)
assert np.allclose(map_computer.recall_per_class[0], 1.0)
assert map_computer.number_true_detection_per_class[0] == 1
assert map_computer.number_false_detection_per_class[0] == 0
assert map_computer.number_found_ground_truth_per_class[0] == 5
assert map_computer. number_missed_ground_truth_per_class[0] == 0
map_computer.reset()
map_computer.update([[5, 2, 15, 9, 1, 0],
[18, 10, 26, 15, 1, 0],
[18, 10, 26, 15, 1, 0],
[18, 10, 26, 15, 1, 0],
[18, 10, 26, 15, 1, 0]], [[0, 0, 26, 26, 0]])
_ = map_computer.compute()
assert np.allclose(map_computer.precision_per_class[0], 1.0)
assert np.allclose(map_computer.recall_per_class[0], 1.0)
assert map_computer.number_true_detection_per_class[0] == 5
assert map_computer.number_false_detection_per_class[0] == 0
assert map_computer.number_found_ground_truth_per_class[0] == 1
assert map_computer.number_missed_ground_truth_per_class[0] == 0
def test_accumulate_pg_coco(self):
map1 = MeanAveragePrecisionMetric(0.01, 'coco')
map1.update(detections[0], gt)
map2 = MeanAveragePrecisionMetric(0.01, 'coco')
map2.update(detections[0], gt)
map2.update(detections[0], gt)
assert np.isclose(map1.compute(), map2.compute())
def test_empty_equals(self):
detections = np.array([[2, 6, 11, 16, 0.9, 0], [20, 11, 45, 25, 0.8, 0]])
ground_truths = np.array([[2, 6, 11, 16, 0], [20, 11, 45, 25, 0]])
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
temp = map_computer.compute()
map_computer.update(np.array([]), np.array([]))
assert map_computer.compute() == temp
def test_empty(self):
detections = np.array([])
ground_truths = np.array([])
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
assert np.isnan(map_computer.compute())
with pytest.warns(RuntimeWarning, match='Mean of empty slice'):
map_computer.compute()
def _ground_truth_label_test(detections, ground_truths, ground_truth_labels_list):
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
assert sorted(list(map_computer.ground_truth_labels)) == sorted(ground_truth_labels_list)
map_computer.compute()
assert sorted(list(map_computer.precision_per_class.keys())) == sorted(ground_truth_labels_list)
assert sorted(list(map_computer.recall_per_class.keys())) == sorted(ground_truth_labels_list)
assert sorted(list(map_computer.average_precision_per_class.keys())) == sorted(ground_truth_labels_list)
assert sorted(list(map_computer.number_false_detection_per_class.keys())) == sorted(ground_truth_labels_list)
assert sorted(list(map_computer.number_found_ground_truth_per_class.keys())) == sorted(ground_truth_labels_list)
assert sorted(list(map_computer.number_missed_ground_truth_per_class.keys())) == \
sorted(ground_truth_labels_list)
assert sorted(list(map_computer.number_true_detection_per_class.keys())) == sorted(ground_truth_labels_list)
def test_empty_detection(self):
detections = np.array([])
ground_truths = np.array([[2, 6, 11, 16, 0], [20, 11, 45, 25, 0]])
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
assert map_computer.compute() == 0.0
assert map_computer.number_true_detection_per_class[0] == 0
assert map_computer.number_false_detection_per_class[0] == 0
assert map_computer.number_found_ground_truth_per_class[0] == 0
assert map_computer.number_missed_ground_truth_per_class[0] == 2
def test_empty_ground_truth(self):
detections = np.array([[2, 6, 11, 16, 0.9, 0], [20, 11, 45, 25, 0.8, 0]])
ground_truths = np.array([])
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
assert map_computer.compute() == 0.0
assert map_computer.number_true_detection_per_class[0] == 0
assert map_computer.number_false_detection_per_class[0] == 2
assert map_computer.number_found_ground_truth_per_class[0] == 0
assert map_computer.number_missed_ground_truth_per_class[0] == 0
assert 0 in map_computer.precision_per_class
assert map_computer.precision_per_class[0] == 0
assert 0 in map_computer.recall_per_class
assert np.isnan(map_computer.recall_per_class[0])
assert 0 in map_computer.average_precision_per_class
assert 0 in map_computer.ground_truth_labels
def test_asymmetrical_empty_ground_truth(self):
detections = np.array([[2, 6, 11, 16, 0.9, 0], [20, 11, 45, 25, 0.8, 1]])
ground_truths = np.array([[2, 6, 11, 16, 0]])
map_computer = MeanAveragePrecisionMetric(0.5, 'coco')
map_computer.update(detections, ground_truths)
assert np.isclose(map_computer.compute(), 0.5)